Dokaz Hipoteza o Spektralnom Radijusu Planarnih Grafova

Journal of Combinatorial Theory, Series B 126 (2017) 137–161
Contents lists available at ScienceDirect
Journal of Combinatorial Theory,

Series B
www.elsevier.com/locate/jctb
Three conjectures in extremal spectral graph theory

Michael Tait a,1 , Josh Tobin b,2
a
Department of Mathematical Sciences, Carnegie Mellon University, United States
b
Department of Mathematics, University of California, San Diego, United States
a r t i c l e i n f o a b s t r a c t
Article history: We prove three conjectures regarding the maximization of

Received 29 August 2016 spectral invariants over certain families of graphs. Our most
Available online 9 May 2017 difficult result is that the join of P2 and Pn−2 is the unique
graph of maximum spectral radius over all planar graphs.
Keywords:
Spectral radius This was conjectured by Boots and Royle in 1991 and
Planar graph independently by Cao and Vince in 1993. Similarly, we prove a
Graph irregularity conjecture of Cvetković and Rowlinson from 1990 stating that
Extremal graph theory the unique outerplanar graph of maximum spectral radius is
the join of a vertex and Pn−1 . Finally, we prove a conjecture
of Aouchiche et al. from 2008 stating that a pineapple graph
is the unique connected graph maximizing the spectral radius
minus the average degree. To prove our theorems, we use the
leading eigenvector of a purported extremal graph to deduce
structural properties about that graph.
© 2017 Elsevier Inc. All rights reserved.
1. Introduction
Questions in extremal graph theory ask to maximize or minimize a graph invariant

over a fixed family of graphs. Perhaps the most well-studied problems in this area are
E-mail addresses: mtait@cmu.edu (M. Tait), rjtobin@math.ucsd.edu (J. Tobin).

1
Some of this research was done while the first author was supported by NSF grant DMS-1606350.
2
Some of this research was done while both authors were partially supported by NSF grant DMS-1362650.
http://dx.doi.org/10.1016/j.jctb.2017.04.006
0095-8956/© 2017 Elsevier Inc. All rights reserved.
138 M. Tait, J. Tobin / Journal of Combinatorial Theory, Series B 126 (2017) 137–161
Turán-type problems, which ask to maximize the number of edges in a graph which does
not contain fixed forbidden subgraphs. Over a century old, a quintessential example of
this kind of result is Mantel’s theorem, which states that Kn/2,n/2 is the unique graph
maximizing the number of edges over all triangle-free graphs. Spectral graph theory seeks
to associate a matrix to a graph and determine graph properties by the eigenvalues and
eigenvectors of that matrix. This paper studies the maximization of spectral invariants
over various families of graphs. We prove three conjectures for n large enough.
Conjecture 1 (Boots–Royle 1991 [8] and independently Cao–Vince 1993 [10]). The planar
graph on n ≥ 9 vertices of maximum spectral radius is P2 + Pn−2 .
Conjecture 2 (Cvetković–Rowlinson 1990 [13]). The outerplanar graph on n vertices of

maximum spectral radius is K1 + Pn−1 .
Conjecture 3 (Aouchiche et al. 2008 [3]). The connected graph on n vertices that maxi-
mizes the spectral radius minus the average degree is a pineapple graph.
In this paper, we prove Conjectures 1, 2, and 3, with the caveat that we must assume n
is large enough in all of our proofs. We note that the Boots–Royle/Vince–Cao conjecture
is not true when n ∈ {7, 8} and thus some bound on n is necessary.
For each theorem, the rough structure of our proof is as follows. A lower bound on the
invariant of interest is given by the conjectured extremal example. Using this information,
we deduce the approximate structure of a (planar, outerplanar, or connected) graph
maximizing this invariant. We then use the leading eigenvalue and eigenvector of the
adjacency matrix of the graph to deduce structural properties of the extremal graph.
Once we know the extremal graph is “close” to the conjectured graph, we show that it
must be exactly the conjectured graph. The majority of the work in each proof is done in
the step of using the leading eigenvalue and eigenvector to deduce structural properties
of the extremal graph.
1.1. History and motivation
Questions in extremal graph theory ask to maximize or minimize a graph invariant

over a fixed family of graphs. This question is deliberately broad, and as such branches
into several areas of mathematics. We already mentioned Mantel’s Theorem as an exam-
ple of a theorem in extremal graph theory. Other classic examples include the following.
Turán’s Theorem [35] seeks to maximize the number of edges over all n-vertex Kr -free
graphs. The Four Color Theorem seeks to maximize the chromatic number over the fam-
ily of planar graphs. Questions about maximum cuts over various families of graphs have
been studied extensively (cf. [2,7,11,18]). The Erdős distinct distance problem seeks to
minimize the number of distinct distances between n points in the plane [16,20].
This paper studies spectral extremal graph theory, the subset of these extremal prob-
lems where invariants are based on the eigenvalues or eigenvectors of a graph. This subset
M. Tait, J. Tobin / Journal of Combinatorial Theory, Series B 126 (2017) 137–161 139
of problems also has a long history of study. Examples include Stanley’s bound maxi-
mizing spectral radius over the class of graphs on m edges [33], the Alon–Bopanna–Serre
Theorem (see [24,30]) and the construction of Ramanujan graphs (see [23]) minimizing
λ2 over the family of d-regular graphs, theorems of Wilf [36] and Hoffman [22] relating
eigenvalues of graphs to their chromatic number, and many other examples. Very re-
cently, Bollobás, Lee, and Letzter studied maximizing the spectral radius of subgraphs
of the hypercube on a fixed number of edges [6].
A bulk of the recent work in spectral extremal graph theory is by Nikiforov, who
has considered maximizing the spectral radius over several families of graphs. Using
the fundamental inequality that λ1 (A(G)) ≥ 2e(G)/n, Nikiforov recovers several classic
results in extremal graph theory. Among these are spectral strengthenings of Turán’s
Theorem [25], the Erdős–Stone–Bollobás Theorem [27], and the Kővari–Sós–Turán The-
orem regarding the Zarankiewicz problem [28] (this was also worked on by Babai and
Guiduli [4]). For many other similar results of Nikiforov, see [29].
We now turn to the history specific to our theorems.
The study of spectral radius of planar graphs has a long history, dating back to
at least Schwenk and Wilson [32]. This direction of research was further motivated by
applications where the spectral radius is used as a measure of the connectivity of a
network, in particular for planar networks in areas such as geography, see for example
[8] and its references. To compare connectivity of networks to a theoretical upper bound,
geographers were interested in finding the planar graph of maximum spectral radius.
To this end, Boots and Royle and independently Cao and Vince conjectured that the
extremal graph is P2 + Pn−2 [8,10]. Several researchers have worked on this problem
and successively improved upon the best theoretical upper bound, including [37,10,38,
19,39,15]. Other related problems have been considered, for example Dvořák and Mohar
found an upper bound on the spectral radius of planar graphs with a given maximum
degree [14]. Work has also been done maximizing the spectral radius of graphs on surfaces
of higher genus [15,38,39]. We would also like to note that it is claimed in [15] that Guiduli
and Hayes proved Conjecture 1 for sufficiently large n. However, this preprint has never
appeared, and the authors could not be reached for comment on it.
Conjecture 2 appears in [13], where the authors mention that it is related to the study
of various subfamilies of Hamiltonian graphs. Rowlinson [31] made partial progress on
this conjecture, which was also worked on by Cao and Vince [10] and Zhou–Lin–Hu [40].
Various measures of graph irregularity have been proposed and studied (cf. [1,5,12,26]
and references therein). These measures capture different aspects of graph irregularity
and are incomparable in general. Because of this, a way to understand which graph
properties each invariant gauges is to look at the extremal graph. For several of the
measures, the graph of maximal irregularity with respect to that measure has been
determined [5,9,21,34]. One such invariant is the spectral radius of the graph minus
its average degree, and Conjecture 3 proposes that the extremal connected graph is a
pineapple graph.
1.2. Notation and preliminaries
Let G be a connected graph and A the adjacency matrix of G. For sets X, Y ⊂ V (G)
we will let e(X) be the number of edges in the subgraph induced by X and e(X, Y )
be the number of edges with one endpoint in X and one endpoint in Y . For a vertex
v ∈ V (G), we will use N (v) to denote the neighborhood of v and dv to denote the degree
of v. For graphs G and H, G + H will denote their join.
Let λ1 ≥ λ2 ≥ · · · ≥ λn be the eigenvalues of A, and let v be an eigenvector cor-
responding to λ1 . By the Perron–Frobenius Theorem, v has all positive entries, and it
will be convenient for us to normalize so that the maximum entry is 1. For a vertex
u ∈ V (G), we will use vu to denote the eigenvector entry of v corresponding to u. With
this notation, for any u ∈ V (G), the eigenvector equation becomes

λ1 vu = vw . (1)
w∼u
Throughout the paper, we will use x to denote the vertex with maximum eigenvector
entry equal to 1. If there are multiple such vertices, choose and fix x arbitrarily among
them. Since x = 1, (1) applied to x becomes

λ1 = vy . (2)
y∼x
Note that this implies λ1 ≤ dx . The next inequality is a simple consequence of

our normalization and an easy double counting argument, but will be used extensively
throughout the paper and warrants special attention. We note that this reasoning has
been used previously, first by Favaron, Mahéo, and Saclé [17]. Multiplying both sides of
(2) by λ1 and applying (1) gives

λ21 = vz = vz + vz
y∼x z∼y y∼x z∼y y∼x z∼y
z∈N (x) z ∈N
/ (x)
≤ 2e (N (x)) + e (N (x), V (G) \ N (x), ) (3)
where the last inequality follows because each eigenvector entry is at most 1, and because
each eigenvector entry appears at the end of a walk of length 2 from x: each edge with
both endpoints in N (x) is the second edge of a walk of length 2 from x exactly twice
and each edge with only one endpoint in N (x) is the second edge of a walk of length 2
from x exactly once.
We will also use the Rayleigh quotient characterization of λ1 :
zt Az
λ1 = max . (4)
z=0 zt z
In particular, this definition of λ1 and the Perron–Frobenius Theorem imply that if

H is a strict subgraph of G, then λ1 (A(G)) > λ1 (A(H)). Another consequence of (4)
that we use frequently is that λ1 ≥ 2m
n , the average degree of G.
1.3. Applying (3)
Our three main results begin by using (3) to deduce structural properties about the
corresponding extremal graphs. To illustrate this technique, in this subsection we use
(3) to give short proofs of two old results. We include this as a quick way for the reader
to become acquainted with our notation and how we will use (3).
Theorem 1 (Mantel’s Theorem). Let G be a triangle-free graph on n vertices. Then G

contains at most n2 /4 edges. Equality occurs if and only if G = Kn/2n/2 .
Proof. If G is triangle-free, then e(N (x)) = 0. Using λ1 ≥ 2m

n and (3) gives
4(e(G))2 n n

≤ e(N (x), V (G) \ N (x)) ≤ .
n2 2 2
Equality may occur only if e(N (x), V (G) \ N (x)) = n2 /4. The only bipartite graph
with this many edges is Kn/2n/2 , and thus Kn/2n/2 is a subgraph of G. But G is
triangle-free, and so G = Kn/2n/2 . 2
We note that one can attempt to use a similar argument to prove Turán’s theorem for
ex(n, Kr ), but because of the presence of the term (e(G))2 , one must use the integrality
of e(G) to deduce the result, and this approach fails when r gets larger than a small
constant.
Theorem 2 (Stanley’s Bound [33]). Let G have m edges. Then
1 √
λ1 ≤ −1 + 1 + 8m .
2
Equality occurs if and only if G is a clique and isolated vertices.
Proof. Using (3) gives

λ21 = vz + 1 ≤ 2(m − dx ) + dx ≤ 2m − λ1 ,
x∼y y∼z x∼y
z=x
where the last inequality holds because λ1 ≤ dx . The result follows by the quadratic
formula. Examining (3) shows that equality holds if and only if E(G) is contained in
the closed neighborhood of x, dx = λ1 , and for each y ∼ x, vy = 1. Since x was chosen
arbitrarily amongst vertices of eigenvector entry 1, any vertex of eigenvector entry 1 must
contain E(G) in its closed neighborhood. Thus G is a clique plus isolated vertices. 2
Fig. 1. The graph P1 + Pn−1 .
1.4. Outline of the paper
Section 3 contains our strongest result, the proof of Conjecture 1. In Section 2 we

prove Conjecture 2 and in Section 4 we prove Conjecture 3.
2. Outerplanar graphs of maximum spectral radius
Let G be a graph. As before, let the first eigenvector of the adjacency matrix of G be
v normalized so that maximum entry is 1. Let x be a vertex with maximum eigenvector
entry, i.e. vx = 1. Throughout let G be an outerplanar graph on n vertices with maximal
adjacency spectral radius. λ1 will refer to λ1 (A(G)).
Two consequences of G being outerplanar that we will use frequently are that G has
at most 2n − 3 edges and G does not contain K2,3 as a subgraph. An outline of our proof
is as follows. We first show that there is a single vertex of large degree and that the
remaining vertices have small eigenvector entry (Lemma 5). We use this to show that
the vertex of large degree must be adjacent to every other vertex (Lemma 6). From here
it is easy to prove that G must be K1 + Pn−1 (Fig. 1).
We begin with an easy lemma that is clearly not optimal, but suffices for our needs.
√
Lemma 3. λ1 > n − 1.
Proof. The star K1,n−1 is outerplanar, and cannot be the maximal outerplanar graph
with respect to spectral radius because it is a strict subgraph of other outerplanar graphs
√
on the same vertex set. Hence, λ1 (G) > λ1 (K1,n ) = n − 1. 2
√
Lemma 4. For any vertex u, we have du > vu n − 11 n.
Proof. Let A be the neighborhood of u, and let B = V (G) \ (A ∪ {u}). We have

λ21 vu = vz ≤ du + vz + z.
y∼u z∼y y∼u z∈N (y)∩A y∼u z∈N (y)∩B
By outerplanarity, each vertex in A has at most two neighbors in A, otherwise G would

contain a K2,3 . In particular,

vz ≤ 2 vy = 2λ1 vu .
y∼u z∈N (y)∩A y∼u
Similarly, each vertex in B has at most 2 neighbors in A. So
2 4e(G) 4(2n − 3)
vz ≤ 2 vz ≤ dz ≤ ≤ ,
y∼u z∈N (y)∩B
λ1 λ1 λ1
z∈B z∈B
as e(G) ≤ 2n − 3 by outerplanarity. So, using Lemma 3 we have

√
vz < 8 n.
y∼u z∈N (y)∩B
Combining the above inequalities yields

√
λ21 vu − 2λ1 vu < du + 8 n.
Again using Lemma 3 we get

√ √ √
vu n − 11 n < (n − 1 − 2 n − 1)vu − 8 n < du . 2
√ √
Lemma 5. We have dx > n − 11 n and for every other vertex u, vu < C1 / n for some
absolute constant C1 , for n sufficiently large.
Proof. The bound on dx follows immediately from the previous lemma and the normal-
ization that vx = 1. Now consider any other vertex u. We know that G contains no K2,3 ,
√ √
so du < 12 n, otherwise u and x share n neighbors, which yields a K2,3 if n ≥ 9. So
√ √
12 n > du > vu n − 11 n,
√
that is, vu < 23/ n. 2
Lemma 6. Let B = V (G) \ (N (x) ∪ {x}). Then

√
vz < C2 / n
z∈B
for some absolute constant C2 .

√
Proof. From the previous lemma, we have |B| < 11 n. Now
1 √ 23
vz ≤ 23/ n dz = √ (e(A, B) + 2e(B)) .
λ1 λ1 n
z∈B z∈B
√
Each vertex in B is adjacent to at most two vertices in A, so e(A, B) ≤ 2|B| < 22 n.
√
The graph induced on B is outerplanar, so e(B) ≤ 2|B| − 3 < 22 n. Finally, using the
√
fact that λ1 > n − 1, we get the required result. 2
Theorem 7. For sufficiently large n, G is the graph K1 + Pn−1 , where + represents the
graph join operation.
Proof. First we show that the set B above is empty, i.e. x is adjacent to every other
vertex. If not, let y ∈ B. Now y is adjacent to at most two vertices in A, and so by
Lemma 5 and Lemma 6,
√ √
vz < vz + 2C1 / n < (C2 + 2C1 )/ n < 1
z∼y z∈B
when n is large enough. Let G+ be the graph obtained from G by deleting all edges
incident to y and replacing them by the single edge {x, y}. The resulting graph is outer-
planar. Then, using the Rayleigh quotient,
vt (A+ − A)v 2vy

λ1 (A ) − λ1 (A) ≥
+
t
= t 1− vz > 0.
vv vv z∼y
This contradicts the maximality of G. Hence B is empty.

Now x is adjacent to every other vertex in G. Hence every vertex other than x has
degree less than or equal to 3. Moreover, the graph induced by V (G) \ {x} cannot
contain any cycles, as then G would not be outerplanar. It follows that G is a subgraph
of K1 + Pn−1 , and maximality ensures that G must be equal to K1 + Pn−1 . 2
3. Planar graphs of maximum spectral radius
As before, let G be a graph with first eigenvector normalized so that maximum entry
is 1, and let x be a vertex with maximum eigenvector entry, i.e. vx = 1. Let m = |E(G)|.
For subsets X, Y ⊂ V (G) we write E(X) for the set of edges induced by X and E(X, Y )
for the set of edges with one endpoint in X and one endpoint in Y . As before, we let
e(X, Y ) = |E(X, Y )|. We will often assume n is large enough without saying so explicitly.
Throughout the section, let G be the planar graph on n vertices with maximum spectral
radius, and let λ1 denote this spectral radius.
We will use frequently that G has no K3,3 as a subgraph, that m ≤ 3n − 6, and that
any bipartite subgraph of G has at most 2n − 4 edges. The outline of our proof is as
follows. We first show that G has two vertices that are adjacent to most of the rest of the
graph (Lemmas 8–11). We then show that the two vertices of large degree are adjacent
(Lemma 13), and that they are adjacent to every other vertex (Lemma 14). The proof
of the theorem follows readily.
√ √
Lemma 8. 6n > λ1 > 2n − 4.
Fig. 2. The graph P2 + Pn−2 .
Proof. For the lower bound, first note that the graph K2,n−2 is planar and is a strict
subgraph of some other planar graphs on the same vertex set (Fig. 2). Since G has
maximum spectral radius among all planar graphs on n vertices,
√
λ1 > λ1 (K2,n−2 ) = 2n − 4.
For the upper bound, since the sum of the squares of the eigenvalues equals twice the
√
number of edges in G, which is at most 6n −12 by planarity, we get that λ1 < 6n − 12 <
√
6n. 2
Next we partition the graph into vertices of small eigenvector entry and those with
large eigenvector entry. Fix > 0, whose exact value will be chosen later. Let
L := {vz ∈ V (G) : vz > }

√
and S = V (G) \ L. For any vertex z, equation (1) gives vz 2n − 4 < vz λ1 ≤ dz .
Therefore,
√
2(3n − 6) ≥ dz ≥ dz ≥ |L| 2n − 4,
z∈V (G) z∈L
√
yielding |L| ≤ 3 2n−4
. Since the subgraph of G consisting of edges with one endpoint
in L and one endpoint in S is a bipartite planar graph, we have e(S, L) ≤ 2n − 4, and
since the√subgraphs induced by S and by L are each planar, we have e(S) ≤ 3n − 6 and
e(L) ≤ 9 2n−4
.
Next we show that there are two vertices adjacent to most of S. The first step towards
this is an upper bound on the sum of eigenvector entries in both L and S.
Lemma 9.
√ 18
vz ≤ 2n − 4 + (5)

z∈L
and
√
vz ≤ (1 + 3) 2n − 4. (6)
z∈S
Proof.
⎛ ⎞
⎜ ⎟
λ1 vz = vy = ⎝ vy + vy ⎠
z∈L z∈L y∼z z∈L y∼z y∼z
y∈S y∈L
√
18 2n − 4
≤ e(S, L) + 2e(L) ≤ (2n − 4) + .

Dividing both sides by λ1 and using Lemma 8 gives (5).
On the other hand,

λ1 vz = vy ≤ 2e(S) + e(S, L) ≤ (6n − 12) + (2n − 4).
z∈S z∈S y∼z
Dividing both sides by λ1 and using Lemma 8 gives (6). 2
Now, for u ∈ L we have

√
vu 2n − 4 ≤ λ1 vu = vy = vy + vy ≤ vy + vy .
y∼u y∼u y∼u y∈L y∼u
y∈L y∈S y∈S
By (5), this gives

√ 18
vy ≥ (vu − ) 2n − 4 − . (7)
y∼u
y∈S
The equations (6) and (7) imply that if u ∈ L and vu is close to 1, then the sum of
the eigenvector entries of vertices in S not adjacent to u is small. The following lemma
is used to show that u is adjacent to most vertices in S.
Lemma 10. For all z we have vz > √1 .

6n
Proof. By way of contradiction assume vz ≤ √16n < λ11 . By equation (1) z cannot be
adjacent to x, since x has eigenvector entry 1. Let H be the graph obtained from G
by removing all edges incident with z and making z adjacent to x. Using the Rayleigh
quotient, we have λ1 (H) > λ1 (G), a contradiction. 2
Now letting u = x and combining (7) and (6), we get
√ √ 18
(1 + 3) 2n − 4 ≥ vy + vy ≥ vy + (1 − ) 2n − 4 − .

y∈S y∈S y∈S
yx y∼x yx
Now applying Lemma 10 gives
1 √ 18
|{y ∈ S : y x}| √ ≤ 4 2n − 4 + .
6n
For n large enough, we have |{y ∈ S : y x}| ≤ 14n. So x is adjacent to most of S.

Our next goal is to show that there is another vertex in L that is adjacent to most of S.
Lemma 11. There is a w ∈ L with w = x such that vw > 1 − 24 and |{y ∈ S : y
w}| ≤ 94n.
Proof. By equation (1), we see

⎛ ⎞ ⎛ ⎞

λ21 = vz ≤ ⎝ vu + vv ⎠ − vy = ⎝ vu + vv ⎠ − λ1 .
y∼x z∼y uv∈E(G) y∼x uv∈E(G)
√
Rearranging and noting that e(S) ≤ 3n − 6 and e(L) ≤ 9 2n−4
since S and L both
induce planar subgraphs gives

2n − 4 ≤ λ21 + λ1 ≤ vu + vv
uv∈E(G)
⎛ ⎞ ⎛ ⎞ ⎛ ⎞

=⎝ vu + vv ⎠ + ⎝ vu + vv ⎠ + ⎝ vu + vv ⎠
uv∈E(S,L) uv∈E(S) uv∈E(L)
⎛ ⎞
√
18 2n − 4
≤⎝ ⎠
vu + vv + (6n − 12) + .

uv∈E(S,L)
So for n large enough,
⎛ ⎞ ⎛ ⎞
⎜ ⎟ ⎜ ⎟
(2 − 7)n ≤ vu + vv = ⎝ vu + vv ⎠ + ⎜
⎝ vu + vv ⎟
⎠
uv∈E(S,L) uv∈E(S,L) uv∈E(S,L)
u=x u=x

≤ e(S, L) + dx + vu ,
uv∈E(S,L)
u=x
giving

vu ≥ (1 − 9)n.
uv∈E(S,L)
u=x
Now since dx ≥ |S| − 14n > (1 − 15)n, and e(S, L) < 2n, the number of terms in
the left hand side of the sum is at most (1 + 15)n. By averaging, there is a w ∈ L such
that
1 − 9
vw ≥ > 1 − 24.
1 + 15
Applying (7) and (6) to this w gives
√ √ 18
(1 + 3) 2n − 4 ≥ vy + vy ≥ vy + (1 − 21) 2n − 4 + ,

y∈S y∈S y∈S
yw y∼w yw
and applying Lemma 10 gives that for n large enough
|{y ∈ S : y w}| ≤ 94n. 2
In the rest of the section, let w be the vertex from Lemma 11. So vx = 1 and vw >
1 − 24, and both are adjacent to most of S. Our next goal is to show that the remaining
vertices are adjacent to both x and w. Let B = N (x) ∩N (w) and A = V (G) \{x ∪w ∪B}.
We show that A is empty in two steps: first we show the eigenvector entries of vertices
in A are as small as we need, which we then use to show that if there is a vertex in A
then G is not extremal.
Lemma 12. Let v ∈ V (G) \ {x, w}. Then vv < 1

10 .
Proof. We first show that the sum over all eigenvector entries in A is small, and then
we show that each eigenvector entry is small. Note that for each v ∈ A, v is adjacent to
at most one of x and w, and is adjacent to at most 2 vertices in B (otherwise G would
contain a K3,3 and would not be planar). Thus

λ1 vv ≤ dv ≤ 3|A| + 2e(A) < 9|A|,
v∈A v∈A
where the
√
last inequality holds by e(A) < 3|A| since A induces a planar graph. Now, since
|L| < 3 2n−4
< n for n large enough, we have |A| ≤ (14 + 94 + 1)n (by Lemma 11).
Therefore
9 · 109 · n
vv ≤ √ .
v∈A
2n − 4
Now any v ∈ V (G) \{x, w} is adjacent to at most 4 vertices in B ∪{x, w}, as otherwise
we would have a K3,3 as above. So we get
√
λ1 vv = vu ≤ 4 + vu ≤ 4 + vu ≤ C n,
u∼v u∼v u∈A
u∈A
where C is an absolute constant not depending on . Dividing both sides by λ1 and

choosing small enough yields the result. 2
We use the fact that the eigenvector entries in A are small to show that if v ∈ A (i.e.
v is not adjacent to both x and w), then removing all edges from v and adding edges
from it to x and w increases the spectral radius, showing that A must be empty. To do
this, we must be able to add edges from a vertex to both x and w and have the resulting
graph remain planar. This is accomplished by the following lemma.
Lemma 13. If G is extremal, then x ∼ w.
Once x ∼ w, one may add a new vertex adjacent to only x and w and the resulting
graph remains planar.
Proof of Lemma 13. From above, we know that for any δ > 0, we may choose small
enough so that when n is sufficiently large we have dx > (1 − δ)n and dw > (1 − δ)n.
By maximality of G, we also know that G has precisely 3n − 6 edges, and by Euler’s
formula, any planar drawing of G has 2n − 4 faces, each of which is bordered by precisely
three edges of G (because in a maximal planar graph, every face is a triangle).
Now we obtain a bound on the number of faces that x and w must be incident to. Let
X be the set of edges incident to x. Each edge in G is incident to precisely two faces, and
each face can be incident to at most two edges in X (again, since each face is a triangle
by maximality). So x is incident to at least |X| = dx ≥ (1 − δ)n faces. Similarly, w is
incident to at least (1 − δ)n faces.
Let F1 be the set of faces that are incident to x, and then let F2 be the set of faces
that are not incident to x, but which share an edge with a face in F1 . Let F = F1 ∪ F2 .
We have |F1 | ≥ (1 − δ)n. Now each face in F1 shares an edge with exactly three other
faces: if two faces shared two edges, then since each face is a triangle both faces must
be bounded by the same three edges; this cannot happen, except in the degenerate case
when n = 3. At most two of these three faces are in F1 , and so |F2 | ≥ |F1 |/3 ≥ (1 −δ)n/3.
Hence, |F | ≥ (1 − δ)4n/3, and so the sum of the number of faces in F and the number of
faces incident to w is larger than 2n − 4. In particular, there must be some face f that
both belongs to F and is incident to w.
Since f ∈ F , then either f is incident to x or f shares an edge with some face that
is incident to x. If f is incident to both x and w, then x is adjacent to w and we are
done. Otherwise, f shares an edge {y, z} with a face f that is incident to x. In this
case, deleting the edge {y, z} and inserting the edge {x, w} yields a planar graph G . By
Lemma 12, the product of the eigenvector entries of y and z is less than 1/100, which
is smaller than the product of the eigenvector entries of x and w. This implies that
λ1 (G ) > λ1 (G), which is a contradiction. 2
We now show that every vertex besides x and w is adjacent to both x and w.
Lemma 14. A is empty.
Proof. Assume that A is nonempty. A induces a planar graph, therefore if A is nonempty,

then there is a v ∈ A such that |N (v) ∩ A| < 6. Further, v has at most 2 neighbors in
B (otherwise G would contain a K3,3 ). Recall that v is the principal eigenvector for the
adjacency matrix of G. Let H be the graph with vertex set V (G) ∪ {v } \ {v} and edge
set E(H) = E(G \ {v}) ∪ {v x, v w}. By Lemma 13, H is a planar graph. Then
vT vλ1 (H)
≥ vT A(H)v

= vT A(G)v − 2 vv vz + 2vv (vw + vx )
z∼v
1
≥ vT A(G)v − 14 · vv · −2 vv vz + 2vv (vw + vx ) (by Lemma 12)
10 z∼v
z∈{w,x}
14
≥ vT A(G)v − vv + 2vv vw (|N (v) ∩ {x, w}| ≤ 1)
10
> vT A(G)v (as vw > 7/10)
T
= v vλ1 (G).
So λ1 (H) > λ1 (G) and H is planar, i.e. G is not extremal, a contradiction. 2
We now have that if G is extremal, then K2 + In−2 , the join of an edge and an
independent set of size n − 2, is a subgraph of G. Finishing the proof is straightforward.
Theorem 15. For n ≥ N0 , the unique planar graph on n vertices with maximum spectral
radius is K2 + Pn−2 .
Proof. By Lemmas 13 and 14, x and w have degree n − 1. We now look at the set
B = V (G) \ {x, w}. For v ∈ B, we have |N (v) ∩ B| ≤ 2, otherwise G contains a copy
of K3,3 . Therefore, the graph induced by B is a disjoint union of paths, cycles, and
isolated vertices. However, if there is some cycle C in the graph induced by B, then
C ∪ {x, w} is a subdivision of K5 . So the graph induced by B is a disjoint union of paths
and isolated vertices. However, if B does not induce a path on n − 2 vertices, then G is
a strict subgraph of K2 + Pn−2 , and we would have λ1 (G) < λ1 (K2 + Pn−2 ). Since G is
extremal, B must induce Pn−2 and so G = K2 + Pn−2 . 2
4. Connected graphs of maximum irregularity
Throughout this section, let G be a graph on n vertices with spectral radius λ1 and
first eigenvector normalized so that vx = 1. Throughout we will use d = 2e(G)/n to
Fig. 3. The pineapple graph, P A(m, n).
denote the average degree. We will also assume that G is the connected graph on n
vertices that maximizes λ1 − d.
To show that G is a pineapple graph (Fig. 3) we first show that λ1 ∼ n2 and d ∼
n n
4 (Lemma 16). Then we show that there exists a vertex with degree close to 2 and
eigenvector entry close to 1 (Lemma 18). We use this to show that there are many
vertices of degree about n2 , that these vertices induce a clique, and further that most of
the remaining vertices have degree 1 (Lemma 19 and Proposition 20). We complete the
proof by showing that all vertices not in the clique have degree 1 and that they are all
adjacent to the same vertex.
We remark that once we show that G is a pineapple graph, the small question remains
of which pineapple graph maximizes λ1 − d. Optimization of a cubic polynomial shows
that G is a pineapple with clique size n2 + 1 (see [3], section 6).
√ 2e(G) √
Lemma 16. We have λ1 (G) = n
2 + c1 n and n = n
4 + c2 n, where |c1 |, |c2 | < 1.
Proof. By eigenvalue interlacing, PA(p, q) has spectral radius at least p − 1. Setting

H = PA n2 + 1, n2 − 1 , we have
2e(H) n 3
λ1 (H) − ≥ − .
n 4 2
On the other hand, an inequality of Hong [37] gives
λ21 ≤ 2e(G) − (n − 1).
It follows that
λ21 1
d≥ +1− . (8)
n n
Setting λ1 = pn and applying (8), we have λ1 − d ≤ pn − p2 n − 1 + n1 . The right hand

side of the inequality is maximized at p = 1/2, giving
n 3 n 1
− ≤ λ1 − d ≤ − 1 + . (9)
4 2 4 n
n √
Next setting λ1 = 2 + c1 n, (8) gives
n √ 1
d≥ + c1 n + c21 + 1 − ,
4 n
whereas (9) implies
n 3 n √ 3
d ≤ λ1 − + = + c1 n + . (10)
4 2 4 2
Together, these imply |c1 | < 1 and prove both statements for n large enough. 2
Lemma 17. There exists a constant c3 not depending on n such that
1 √
0≤ dy − λ1 vy ≤ c3 n.
|N (x)| y∼x
Proof. From the inequality of Hong,

λ1 vy = λ21 ≤ dn − (n − 1).
y∼x
Rearranging and applying Lemma 16, we have

0≤ (dy − λ1 vy ) = O n3/2 .
y∼x
By equation (1) again, and because the first eigenvector is normalized with vx = 1, we
have

λ1 = vy ≤ dx ,
y∼x
giving dx = Ω(n). Combining, we have
1 √
(dy − λ1 vy ) = O n ,
|N (x)| y∼x
where the implied constant is independent of n. 2
Now we fix a constant > 0, whose exact value will be chosen later. The next lemma
implies that close to half of the vertices of G have eigenvector entry close to 1 for n
sufficiently large, depending on the chosen . We follow that with a proposition which
outlines the approximate structure of G, and then finally use variational arguments to
deduce that G is exactly a pineapple graph.
√
Lemma 18. There exists a vertex u = x with vu > 1 − 2 and du − λ1 vu = O( n).
Moreover du ≥ (1/2 − 2) n.
Proof. We proceed by first showing a weaker result: that there is a vertex y with vy >
√
2 − and dy − λ1 vy = O( n), and additionally that y ∈ N (x). We will then use this to
1
obtain the required result.

Let A := {z ∼ x : vz > 12 − }. By Lemma 16,
n √
λ1 = + c1 n,
2
where |c1 | < 1. Since 0 < vz ≤ 1 for all z ∼ x, we see that |A| ≥ δ n where δ is a
√
positive constant that depends only on . Let B = {z ∼ x : dz − λ1 vz > K n}, where
K is a fixed constant whose exact value will be chosen later. Now
1 1 1 √
(dy − λ1 vy ) ≥ (dz − λ1 vz ) ≥ |B|K n.
|N (x)| y∼x |N (x)| n
z∈B
By Lemma 17, |B| ≤ cK3 n. Therefore, for K large enough depending only on , we have
|A ∩ B c | > 0. This proves the existence of the vertex y, with the properties claimed at
the beginning of the proof.

Next, we show that there exists a set U ⊂ N (y) such that |U | ≥ 14 − 2 n and
vu ≥ 1 − 2 for all u ∈ U . By Lemma 16,
n
√ 1
+ c1 n − ≤ λ1 vy ≤ dy ,
2 2
1
where |c1 | < 1. So dy ≥ 4 − n for n large enough. Now let C = {z ∼ y : vz < 1 − 2}.
Then
√
K n ≥ dy − λ1 vy = (1 − vz ) ≥ (1 − vz ) ≥ 2|C|.
z∼y z∈C
Therefore
√
1 K n
|N (y) \ C| ≥ − n− .
4 2

Setting U = N (y) \ C, we have |U | > 14 − 2 n for n large enough.
Set D = U ∩ N (x). We will first find a lower bound on |D|. We have

λ21 ≤ dy ≤ 2m − dy .
y∼x y ∈N
/ (x)
Rearranging this we get

λ21 1
d− ≥ dy .
n n
y ∈N
/ (x)
Now applying the bound on d from equation (10) and expression for λ1 in Lemma 16
yields
n √ 2
n √ 3 + c1 n 1
+ c1 n + − 2
≥ dy ,
4 2 n n
y ∈N
/ (x)
which implies that

3 3
n≥ − c1 n ≥
2
dy ≥ dy ≥ |U \ N (x)|(1 − 2)λ1 .
2 2
y ∈N
/ (x) y∈U \N (x)
So
3 n 3 1
|U \ N (x)| ≤ = .
2(1 − 2) λ1 2(1 − 2) 1/2 + c1 n−1/2
In particular, |D| ≥ ( 14 − c )n.

Now by the same argument used at the start of the proof to show the existence of the
√
vertex y, we have some vertex u ∈ D with du − λ1 vu = O( n). Finally
√
du ≥ vu λ1 ≥ (1 − 2)(n/2 + c1 n) ≥ (1/2 − 2) n. 2
Lemma 19. Let x, y be two vertices in G. If vx vy > 1/2 + n−1/2 + 5n−1 , then x and y
are adjacent. On the other hand, if vx vy < 1/2 − 3 then x and y are not adjacent.
Proof. We begin by bounding the dot product of the leading eigenvector v with itself.
We will show that
n √ n √
+ n + 5 ≥ vt v > − 2n − O( n). (11)
2 2
First, we show the lower bound. With u from the previous lemma, by Cauchy–Schwarz
we have

2
1 (λ1 vu )2
vv≥
t
v2z ≥ vz = .
z∼u
du z∼u
du
By Lemma 18, we then have

√
(du − O( n))2 √ n √
vv≥
t
≥ du − O( n) > − 2n − O( n).
du 2
C
For the upper bound of inequality (11), first set E = (N (x) ∪ {x}) . Then
1
vt v = v2z ≤ vz ≤ 1 + vz + vz ≤ 1 + λ1 + dz .
λ1
z∈V (G) z∈V (G) z∈N (x) z∈E z∈E
From the proof of Lemma 18 we have the bound

3
dz ≤ n.
2
z∈E
Hence
n √ 3 1 n √
vt v ≤ 1 + + c1 n + · ≤ + n + 5.
2 2 1/2 + c1 n−1/2 2
This completes the proof of inequality (11).

1 be the leading eigenvalue of the graph formed by adding the edge {x, y} to G.
Let λ+
Then by (4) we have
vt (A+ − A)v 2vx vy 2vx vy 2vx vy

1 − λ1 ≥
λ+ ≥ ≥ √ = .
t
vv t
vv n/2 + n + 5 n(1/2 + n−1/2 + 5n−1 )
If vx vy > 1/2 + n−1/2 + 5n−1 , then
2 2
1 − d ) − (λ − d) >
(λ+ − = 0.
+
n n
Hence {x, y} must already have been an edge, otherwise this would contradict the max-
imality of G.
Similarly if λ−
1 is the leading eigenvalue of the graph obtained from G by deleting the
edge {x, y}, then
vt (A − A− )v 2vx vy 2vx vy
λ1 − λ−
1 ≤ ≤ √ ≤ ,
t
vv n/2 − 2n − O( n) (1/2 − 3)n
when n is large enough. Now if vx vy < 1/2 − 3, then
(λ1 − d) − (λ− −
1 − d ) < 0. 2
Proposition 20. For n sufficiently large, we can partition the vertices of G into three sets
U, V, W (see Fig. 4) where
(i) vertices in V have eigenvector entry smaller than (2 + )/n and have degree one,
(ii) vertices in U induce a clique, all have eigenvector entry larger than 1 − 20, and
(1/2 − 3)n ≤ |U | ≤ (1/2 + )n,
Fig. 4. Structure of G in Proposition 20. The number beside each set indicates the values of eigenvector
entries within the set. U induces a complete graph and V , W are independent sets. Each vertex in V is
adjacent to exactly one vertex in U , and each vertex in W is adjacent to multiple vertices in U .
(iii) vertices in W have eigenvector entry in the range [1/2 − 4, 1/2 + 21] and are
adjacent only to vertices in U .
Proof. By Lemma 19, any two vertices in G with eigenvector entry 1 are adjacent.
Moreover, it is easy to see that every vertex in G is incident to at least one vertex with
eigenvector entry 1: if not, for each vertex not incident to a vertex with eigenvector
entry 1, delete one of its edges and add a new edge from that vertex to a vertex with
eigenvector entry 1 (such as the vertex x). The resulting graph is connected, will have
the same number of edges as the original graph, and will have strictly larger λ1 (this
can be seen by considering the Rayleigh quotient, as in the proof of Lemma 19). So by
maximality of G, there are no such vertices. This implies that the set of edges that are
incident to a vertex with eigenvector entry 1 spans the vertex set of G. In particular, if
we remove any edge that is not incident to a vertex with eigenvector entry 1, we do not
disconnect the graph. We will use this fact repeatedly in this proof.
(i) Let V consist of all vertices in G with eigenvector entry less than 1/2 − 4. By
Lemma 19, removing any edge incident to a vertex in V strictly increases λ1 − d,
so each vertex in V has degree one. By equation (1), the eigenvector entry of any
such vertex is at most 1/λ1 < (2 + )/n, when n is large enough.
√
(ii) From Lemma 18, we have a vertex u such that du − λ1 vu = O( n). Let X be the
set of neighbors z of u such that vz < 9/10. Then we have
√
(1 − 9/10)|X| ≤ 1 − vy = du − λ1 vu = O( n).
y∼u
√
Hence |X| = O( n). Let U be all vertices in G with eigenvector entry at least 9/10.
So, by Lemma 18
√
|U | ≥ du − |X| ≥ n/2 − 2n − O( n).
For n large enough, we have |U | ≥ (1/2 −3)n. For sufficiently large n, by Lemma 19
these vertices are all adjacent to each other. For the upper bound on |U | we use the
expression for e(G) in Lemma 16
n2 √
|U |(|U | − 1) ≤ 2e(G) ≤ + c2 n n,
4
which implies |U | ≤ (1/2 + )n for large enough n.

Now take any vertex y ∈ U . If x is a vertex with largest eigenvector entry, then

λ1 − λ1 vy ≤ vz ≤ vy + vz . (12)
z∈N (x)\N (y) z∈U C
We have

λ1 vz ≤ dz ≤ 2e(G) − 2|E(U, U )|
z∈U C z∈U C
n2 √
≤ + c2 n n − (1/2 − 3)(1/2 − 3 − 1/n)n2
4
≤ 4n2 ,
for n sufficiently large, where we are using the expression for e(G) given by
Lemma 16. In particular,

vz ≤ 9n.
z∈U C
Finally, by equation (12) we have
1 vy
vy ≥ 1 − vz − ≥ (1 − 20).
λ1 C
λ1
z∈U
(iii) Let W consist of all remaining vertices of G. If a vertex has eigenvector entry smaller
than 1/2 − 4 then it is in V by construction. If a vertex z ∈ W has eigenvector
entry larger than 1/2 + 21 then we have
(1/2 + 21)(1 − 20) > 1/2 + ,
if < 1/50, say. So for sufficiently large n, by Lemma 19 we have that z is adjacent
to every vertex in U . But by the proof of part (ii), this implies that vz > 1 − 20,
which contradicts z ∈ W .
For z ∈ W and any vertex y ∈ U C , then vy vz ≤ (1/2 + 21)(1/2 + 21) < 1/4 + 22
and so by Lemma 19 there is no edge between y and z in the maximal graph G. 2
Theorem 21. For sufficiently large n, G is a pineapple graph.
Proof. Take U, V, W as in the previous lemma. We begin by showing that the set W
must be empty. Proceeding by contradiction, let z be in W . Furthermore let G+ be
the graph obtained by adding edges from z to every vertex in U . We will show that
λ1 (G+ ) − d(G+ ) > λ1 (G) − d(G), which contradicts the maximality of G.
Since the vertex z is adjacent only to vertices in U , and the fact that vertices in U
have eigenvector entry between 1 − 20 and 1, equation (1) yields
λ1 vz
λ1 (1/2 − 4) ≤ λ1 vz ≤ dz (G) ≤ = (1/2 + O())λ1 .
1 − 20
Using the expression for λ1 in Lemma 16, for large enough n we have
n n
(1 − ) ≤ dz (G) ≤ (1 + ) .
4 4
So we can bound the change in the average degrees
2(|U | − (1 − )n/4)
d(G+ ) − d(G) ≤ < 1/2 + 3.
n
Next we find a lower bound on λ1 (G+ ) − λ1 (G). Let w be the vector that is equal to v
on all vertices except z, and equal to 1 for z. Then,
w t A+ w
λ1 (G+ ) ≥ .
wt w
We first find a lower bound for the numerator (with abuse of big-O notation with in-
equalities)
wt A+ w ≥ wt Aw + 2(|U | − dz (G))(1 − O()) ≥ wt Aw + (1/2 − O())n

≥ vt Av + 2dz (G) (1 − vz ) (1 − 20) + (1/2 − O())n
≥ vt Av + 2dz (G) (1/2 − 31) + (1/2 − O())n
≥ vt Av + (3/4 − O())n.
Similarly, we find an upper bound for the denominator
wt w = vt v + 1 − v2z
≤ vt v + 1 − (1/2 − 4)2
≤ vt v + 3/4 + 4.
Combining these, and using the bound on vt v from the proof of Lemma 19, we get
wt A+ w vt Av
λ1 (G+ ) − λ1 (G) ≥ − t
wt w vv
v v(3/4 − O())n − vt Av(3/4 + 4)
t
≥
vt v(vt v + 3/4 + 4)
(3/4 − O())n − (3/4 + 4)λ1 (G)
≥
vt v + 3/4 + 4
= 3/4 + O().
Hence λ1 (G+ ) − λ1 (G) > d(G+ ) − d(G), and by maximality of G we conclude that
W = ∅.
At this point we know that G consists of a clique together with a set of pendant
vertices V . All that remains is to show that all of the pendant vertices are incident to
the same vertex in the clique. Let V = {v1 , v2 , · · · , vk }, and let ui be the unique vertex
in U that vi is adjacent to. Let G+ be the graph obtained from G by deleting the edges
{vi , ui } and adding the edges {vi , x}, where x is a vertex with eigenvector entry 1. Now,
d(G+ ) = d(G), and
vt A+ v vt Av
λ1 (G+ ) − λ1 (G) ≥ − t ,
vt v vv
with equality if and only if v is a leading eigenvector for A+ . We have

k
vt A+ v vt Av 1
− t = t 1 − vui ≥ 0,
vt v vv vv i=1
with equality if and only if vui = 1 for all 1 ≤ i ≤ k. By maximality of G, we have

equality in both of the above inequalities, and so v is a leading eigenvector for G+ , and
every vertex in U incident to a vertex in V has eigenvector entry 1. G+ is a pineapple
graph, and it is easy to see that there is a single vertex in a pineapple graph with
maximum eigenvector entry. It follows that the vertices in V are all adjacent to the same
vertex in U , and hence G is a pineapple graph. 2
Acknowledgments
We would like to thank Vlado Nikiforov for helpful comments.
References
[1] Michael O. Albertson, The irregularity of a graph, Ars Combin. 46 (1997) 219–225.
[2] Noga Alon, Bipartite subgraphs, Combinatorica 16 (3) (1996) 301–311.
[3] Mustapha Aouchiche, Francis K. Bell, Dragiša Cvetković, Pierre Hansen, Peter Rowlinson, Slo-
bodan K. Simić, Dragan Stevanović, Variable neighborhood search for extremal graphs. 16. Some
conjectures related to the largest eigenvalue of a graph, European J. Oper. Res. 191 (3) (2008)
661–676.
[4] László Babai, Barry Guiduli, Spectral extrema for graphs: the Zarankiewicz problem, Electron. J.
Combin. 16 (1) (2009) R123.
[5] Francis K. Bell, A note on the irregularity of graphs, Linear Algebra Appl. 161 (1992) 45–54.
[6] Béla Bollobás, Jonathan Lee, Shoham Letzter, Eigenvalues of subgraphs of the cube, preprint,
arXiv:1605.06360, 2016.
[7] Béla Bollobás, Alex D. Scott, Better bounds for max cut, Contemporary combinatorics 10 (2002)
185–246.
[8] Barry N. Boots, Gordon F. Royle, A conjecture on the maximum value of the principal eigenvalue
of a planar graph, Geogr. Anal. 23 (3) (1991) 276–282.
[9] Graham Brightwell, Peter Winkler, Maximum hitting time for random walks on graphs, Random
Structures Algorithms 1 (3) (1990) 263–276.
[10] Dasong Cao, Andrew Vince, The spectral radius of a planar graph, Linear Algebra Appl. 187 (1993)
251–257.
[11] Fan RK Chung, Spectral Graph Theory, vol. 92, American Mathematical Society, 1997.
[12] Sebastian M. Cioaba, David A. Gregory, Principal eigenvectors of irregular graphs, Electron. J.
Linear Algebra 16 (2007) 366–379.
[13] Dragiša Cvetković, Peter Rowlinson, The largest eigenvalue of a graph: a survey, Linear Multilinear
Algebra 28 (1–2) (1990) 3–33.
[14] Zdeněk Dvořák, Bojan Mohar, Spectral radius of finite and infinite planar graphs and of graphs of
bounded genus, J. Combin. Theory Ser. B 100 (6) (2010) 729–739.
[15] Mark N. Ellingham, Xiaoya Zha, The spectral radius of graphs on surfaces, J. Combin. Theory Ser.
B 78 (1) (2000) 45–56.
[16] Paul Erdős, On sets of distances of n points, Amer. Math. Monthly 53 (5) (1946) 248–250.
[17] Odile Favaron, Maryvonne Mahéo, J-F. Saclé, Some eigenvalue properties in graphs (conjectures of
Graffiti II), Discrete Math. 111 (1) (1993) 197–220.
[18] Michel X. Goemans, David P. Williamson, Improved approximation algorithms for maximum cut
and satisfiability problems using semidefinite programming, J. ACM 42 (6) (1995) 1115–1145.
[19] Barry D. Guiduli, Spectral Extrema for Graphs, PhD thesis, University of Chicago, 1996.
[20] Larry Guth, Nets Hawk Katz, On the Erdős distinct distances problem in the plane, Ann. of Math.
181 (1) (2015) 155–190.
[21] Pierre Hansen, Hadrien Mélot, Groupe d’études et de recherche en analyse des décisions, Variable
Neighborhood Search for Extremal Graphs 9: Bounding the Irregularity of a Graph, Groupe d’études
et de recherche en analyse des décisions, Montréal, 2002.
[22] Alan J. Hoffman, On eigenvalues and colorings of graphs, in: Graph Theory and its Applications,
Proc. Advanced Sem., Math. Research Center, Univ. of Wisconsin, Madison, Wis., 1969, Academic
Press, New York, 1970.
[23] Alexander Lubotzky, Ralph Phillips, Peter Sarnak, Ramanujan graphs, Combinatorica 8 (3) (1988)
261–277.
[24] M. Ram Murty, Ramanujan graphs, J. Ramanujan Math. Soc. 18 (1) (2003) 33–52.
[25] Vladimir Nikiforov, Some inequalities for the largest eigenvalue of a graph, Combin. Probab. Com-
put. 11 (02) (2002) 179–189.
[26] Vladimir Nikiforov, Eigenvalues and degree deviation in graphs, Linear Algebra Appl. 414 (1) (2006)
347–360.
[27] Vladimir Nikiforov, A spectral Erdős–Stone–Bollobás Theorem, Combin. Probab. Comput. 18 (03)
(2009) 455–458.
[28] Vladimir Nikiforov, A contribution to the Zarankiewicz problem, Linear Algebra Appl. 432 (6)
(2010) 1405–1411.
[29] Vladimir Nikiforov, Some new results in extremal graph theory, in: Surveys in Combinatorics,
Cambridge University Press, 2011, pp. 213–218.
[30] A. Nilli, On the second eigenvalue of a graph, Discrete Math. 91 (2) (1991) 207–210.
[31] Peter Rowlinson, On the index of certain outerplanar graphs, Ars Combin. 29 (1990) 221–225.
[32] Allen J. Schwenk, Robin J. Wilson, On the eigenvalues of a graph, in: Lowell W. Beineke, Robin
J. Wilson (Eds.), Selected Topics in Graph Theory, Academic Press, London, 1978, pp. 307–336,
chapter 11.
[33] Richard P. Stanley, A bound on the spectral radius of graphs with e edges, Linear Algebra Appl.
87 (1987) 267–269.
[34] Michael Tait, Josh Tobin, Characterizing graphs of maximum principal ratio, preprint, arXiv:
1511.06378, 2015.
[35] Paul Turán, On an extremal problem in graph theory, Mat. Fiz. Lapok 48 (137) (1941) 436–452.
[36] Herbert S. Wilf, Spectral bounds for the clique and independence numbers of graphs, J. Combin.
Theory Ser. B 40 (1) (1986) 113–117.
[37] Hong Yuan, A bound on the spectral radius of graphs, Linear Algebra Appl. 108 (1988) 135–139.
[38] Hong Yuan, On the spectral radius and the genus of graphs, J. Combin. Theory Ser. B 65 (2) (1995)
262–268.
[39] Hong Yuan, Upper bounds of the spectral radius of graphs in terms of genus, J. Combin. Theory
Ser. B 74 (2) (1998) 153–159.
[40] Jian Zhou, Cuiqin Lin, Guanzhang Hu, Spectral radius of Hamiltonian planar graphs and outerpla-
nar graphs, Tsinghua Sci. Technol. 6 (4) (2001) 350–354.

Dokaz Hipoteza o Spektralnom Radijusu Planarnih Grafova

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Dokaz Hipoteza o Spektralnom Radijusu Planarnih Grafova

Hochgeladen von

Copyright:

Verfügbare Formate

Journal of Combinatorial Theory, Series B 126 (2017) 137–161

Contents lists available at ScienceDirect

Journal of Combinatorial Theory,

Three conjectures in extremal spectral graph theory

Article history: We prove three conjectures regarding the maximization of

Questions in extremal graph theory ask to maximize or minimize a graph invariant

E-mail addresses: mtait@cmu.edu (M. Tait), rjtobin@math.ucsd.edu (J. Tobin).

Conjecture 2 (Cvetković–Rowlinson 1990 [13]). The outerplanar graph on n vertices of

1.1. History and motivation

Questions in extremal graph theory ask to maximize or minimize a graph invariant

1.2. Notation and preliminaries

Note that this implies λ1 ≤ dx . The next inequality is a simple consequence of

≤ 2e (N (x)) + e (N (x), V (G) \ N (x), ) (3)

In particular, this deﬁnition of λ1 and the Perron–Frobenius Theorem imply that if

1.3. Applying (3)

Theorem 1 (Mantel’s Theorem). Let G be a triangle-free graph on n vertices. Then G

Proof. If G is triangle-free, then e(N (x)) = 0. Using λ1 ≥ 2m

4(e(G))2 n n

Theorem 2 (Stanley’s Bound [33]). Let G have m edges. Then

Proof. Using (3) gives

Fig. 1. The graph P1 + Pn−1 .

1.4. Outline of the paper

Section 3 contains our strongest result, the proof of Conjecture 1. In Section 2 we

2. Outerplanar graphs of maximum spectral radius

Proof. Let A be the neighborhood of u, and let B = V (G) \ (A ∪ {u}). We have

By outerplanarity, each vertex in A has at most two neighbors in A, otherwise G would

Similarly, each vertex in B has at most 2 neighbors in A. So

as e(G) ≤ 2n − 3 by outerplanarity. So, using Lemma 3 we have

Combining the above inequalities yields

Again using Lemma 3 we get

Lemma 6. Let B = V (G) \ (N (x) ∪ {x}). Then

for some absolute constant C2 .

vt (A+ − A)v 2vy

This contradicts the maximality of G. Hence B is empty.

3. Planar graphs of maximum spectral radius

Fig. 2. The graph P2 + Pn−2 .

L := {vz ∈ V (G) : vz > }

Dividing both sides by λ1 and using Lemma 8 gives (6). 2

Now, for u ∈ L we have

By (5), this gives

Lemma 10. For all z we have vz > √1 .

Now letting u = x and combining (7) and (6), we get

Now applying Lemma 10 gives

For n large enough, we have |{y ∈ S : y x}| ≤ 14n. So x is adjacent to most of S.

Proof. By equation (1), we see

So for n large enough,

Applying (7) and (6) to this w gives

and applying Lemma 10 gives that for n large enough

|{y ∈ S : y w}| ≤ 94n. 2

Lemma 12. Let v ∈ V (G) \ {x, w}. Then vv < 1

where C is an absolute constant not depending on . Dividing both sides by λ1 and

Lemma 13. If G is extremal, then x ∼ w.

Lemma 14. A is empty.

Proof. Assume that A is nonempty. A induces a planar graph, therefore if A is nonempty,

So λ1 (H) > λ1 (G) and H is planar, i.e. G is not extremal, a contradiction. 2

4. Connected graphs of maximum irregularity

Fig. 3. The pineapple graph, P A(m, n).

Proof. By eigenvalue interlacing, PA(p, q) has spectral radius at least p − 1. Setting

λ21 ≤ 2e(G) − (n − 1).

Setting λ1 = pn and applying (8), we have λ1 − d ≤ pn − p2 n − 1 + n1 . The right hand

whereas (9) implies

Lemma 17. There exists a constant c3 not depending on n such that

Proof. From the inequality of Hong,

Rearranging and applying Lemma 16, we have

4(e(G))2 n n